Image classification method based on linear self-attention Transform

Abstract

The invention relates to the technical field of computer vision, in particular to an image classification method based on a linear self-attention Transform. The method comprises the following steps: S1, sending a picture to an overlapping convolution coding module of a first stage, and coding the picture into a picture token by using roll operation; s2, sending the picture token into a Transform module in the stage, and extracting a picture feature vector; s3, the extracted picture feature vectors are sent to an overlapping convolution coding module of the next stage, and the feature vector dimensionality is increased while the number of the feature vectors is reduced; s4, repeating the step S2 and the step S3, and obtaining a final output vector from the Transform module of the last stage; and S5, converting the final output vector into probability representation through a classifier unit, and completing image classification. According to the method, the image features can be effectively extracted from the image, the calculation complexity of the Transform module is remarkably reduced, and the ability of the model to extract the image features is improved through the overlapping convolutional coding module and the convolutional feedforward neural network module.

Description

technical field [0001] The invention relates to the technical field of computer vision, in particular to an image classification method based on linear self-attention Transformer. Background technique [0002] From the revolutionary performance of AlexNet in the ImageNet classification challenge, the CNN network architecture has developed rapidly. Since then, deeper and more efficient convolutional network structures have been proposed to further promote the wave of deep learning in the field of computer vision, such as VGG, GoogleNet, Resnet, DenseNet, HRNet, and EfficientNet. CNN and its variants have become the main backbone architecture for computer vision applications. [0003] Transformer was first proposed by the Google team in 2017 for translation tasks in the field of natural language processing (NLP). It can improve computational efficiency by using a multi-head attention mechanism to model long-range dependencies and parallel computing. Therefore, the Transform...

AI Technical Summary

Problems solved by technology

Therefore, when the input image resolution is too high, the amount of calculation required for image classification using ViT will be very large.

Moreover, ViT lacks inductive biases such as translation invariance and locality compared to convolutional neural networks, so the amount of data required for ViT model training is greater than the amount of data required for convolutional neural networks.

Images